Hand-off for Claude Code / Grok
Full write-up (file-level, prioritized):
→ docs/PREMORTEM.md
Also indexed in docs/README.md.
TL;DR — fix P0 before wiring agents on ask().action
Until then prefer completionCheck().verdict (+ result_is_echo), not ask().action.
P0
- Noul polarity —
confidenceOf in src/policy.ts uses Math.max(noul, 1-noul), so confident false bands as accept.
- Aggregate lie —
all_parts_at_least_0.7 parses the floor then ignores it (resolveAggregate → plain minConfidence).
- Silent ledger loss —
fileLedger.append empty catch; partial JSONL on mid-flight failure.
- Echo beside verdict — hosts checking only
verdict === "done" can ship restatements.
P1
- Live
timeout_ms: 750 vs eval 30s; judge no-retry vs eval 5×; 30/30 not in npm test; README MIT vs package.json UNLICENSED.
Suggested patch order
See bottom of docs/PREMORTEM.md.
Reviewed 2026-09-21 against src/, policy.json, tests, DESIGN/USAGE/README claims.
Hand-off for Claude Code / Grok
Full write-up (file-level, prioritized):
→
docs/PREMORTEM.mdAlso indexed in
docs/README.md.TL;DR — fix P0 before wiring agents on
ask().actionUntil then prefer
completionCheck().verdict(+result_is_echo), notask().action.P0
confidenceOfinsrc/policy.tsusesMath.max(noul, 1-noul), so confident false bands asaccept.all_parts_at_least_0.7parses the floor then ignores it (resolveAggregate→ plainminConfidence).fileLedger.appendemptycatch; partial JSONL on mid-flight failure.verdict === "done"can ship restatements.P1
timeout_ms: 750vs eval 30s; judge no-retry vs eval 5×; 30/30 not innpm test; README MIT vspackage.jsonUNLICENSED.Suggested patch order
See bottom of
docs/PREMORTEM.md.Reviewed 2026-09-21 against
src/,policy.json, tests, DESIGN/USAGE/README claims.